Robust and exploratory analysis of active mesoscale tectonic zones in Japan...

27
Robust and exploratory analysis of active mesoscale tectonic zones in Japan utilizing the nationwide GPS array Yuzo Toya T , Minoru Kasahara Institute of Seismology and Volcanology, Hokkaido University, N10W8 Sapporo, Hokkaido 060-0810, Japan Received 5 April 2004; accepted 3 February 2005 Available online 29 March 2005 Abstract A monitoring GPS array recently developed in Japan can yield nationwide maps of active inland tectonic zones (ATZs) on a mesoscale, approximately 70 to several hundred kilometers in lateral extent. But it has been difficult to characterize ATZs in Japan, as they are in fact operational on multiple scales and our efforts are often hindered by various irregularities in the data. The key to overcoming these problems would be to gain an insight into the available data before any precise kinematic modeling is performed with indefinite assumptions. In this study, horizontal velocity fields, deduced from the nationwide GPS array, were treated with a set of techniques in robust smoothing and exploratory data analysis that brought out exceptionally powerful mesoscale ATZs, and made them easier to characterize. The resolved ATZs were then retrospectively monitored to study their regional and temporal variations, using a set of approx. 840 observation stations, about 30 km apart, for a 4-year series of fixed observation time-intervals, 810 days each. The smoothing operation involved three steps: (1) imputation of the velocity fields for the purpose of anti-aliasing, (2) robust smoothing of the velocity fields with the median operative, and (3) visualization of deformation-rate distributions in several coordinate independent parameters, and post-filtering. The geometrical resolvability of mesoscale ATZs was confirmed by calibrating the smoothing scheme against synthetic tectonic boundary models before it was applied to the case study in Japan. ATZs in Japan, which are essentially visible as systematic deviations in the velocity fields on the International Terrestrial Reference Frame (ITRF) and as strain rate anomalies, were highlighted sharply along some known tectonic zones, chains of active volcanoes, and areas above low seismic velocity anomalies in the crust and upper mantle, all of which generally paralleled the offshore trench axes. The geometrical agreements among the mapped ATZs and the physical anomalies in the crust are presumably due to their common structural weakness on the mesoscale. In the four main islands of Japan, all but 30–40% of the strain rate anomalies persisted during the entire 6 years of the case study period, while the rest sporadically appeared or disappeared in a period from several months to a few years. The transient shifts in the deformation rates were remarkably synchronous with some nearby major tectonic episodes: large earthquakes and slow events. Differential plate coupling strengths along the subduction zones can also be inferred from the persistent pattern of rotational strain rate anomalies forming clockwise and counterclockwise pairs along the Pacific. Our empirical observations suggest that the first-order features of interseismic crustal deformations in Japan can be characterized as collateral processes behaving in 0040-1951/$ - see front matter D 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.tecto.2005.02.003 T Corresponding author. Tel.: +81 11 706 2643; fax: +81 11 746 7404. E-mail address: [email protected] (Y. Toya). Tectonophysics 400 (2005) 27– 53 www.elsevier.com/locate/tecto

Transcript of Robust and exploratory analysis of active mesoscale tectonic zones in Japan...

  • www.elsevier.com/locate/tecto

    Tectonophysics 400

    Robust and exploratory analysis of active mesoscale tectonic

    zones in Japan utilizing the nationwide GPS array

    Yuzo ToyaT, Minoru Kasahara

    Institute of Seismology and Volcanology, Hokkaido University, N10W8 Sapporo, Hokkaido 060-0810, Japan

    Received 5 April 2004; accepted 3 February 2005

    Available online 29 March 2005

    Abstract

    A monitoring GPS array recently developed in Japan can yield nationwide maps of active inland tectonic zones (ATZs) on a

    mesoscale, approximately 70 to several hundred kilometers in lateral extent. But it has been difficult to characterize ATZs in

    Japan, as they are in fact operational on multiple scales and our efforts are often hindered by various irregularities in the data.

    The key to overcoming these problems would be to gain an insight into the available data before any precise kinematic

    modeling is performed with indefinite assumptions. In this study, horizontal velocity fields, deduced from the nationwide GPS

    array, were treated with a set of techniques in robust smoothing and exploratory data analysis that brought out exceptionally

    powerful mesoscale ATZs, and made them easier to characterize. The resolved ATZs were then retrospectively monitored to

    study their regional and temporal variations, using a set of approx. 840 observation stations, about 30 km apart, for a 4-year

    series of fixed observation time-intervals, 810 days each. The smoothing operation involved three steps: (1) imputation of the

    velocity fields for the purpose of anti-aliasing, (2) robust smoothing of the velocity fields with the median operative, and (3)

    visualization of deformation-rate distributions in several coordinate independent parameters, and post-filtering. The geometrical

    resolvability of mesoscale ATZs was confirmed by calibrating the smoothing scheme against synthetic tectonic boundary

    models before it was applied to the case study in Japan. ATZs in Japan, which are essentially visible as systematic deviations in

    the velocity fields on the International Terrestrial Reference Frame (ITRF) and as strain rate anomalies, were highlighted sharply

    along some known tectonic zones, chains of active volcanoes, and areas above low seismic velocity anomalies in the crust and

    upper mantle, all of which generally paralleled the offshore trench axes. The geometrical agreements among the mapped ATZs

    and the physical anomalies in the crust are presumably due to their common structural weakness on the mesoscale. In the four

    main islands of Japan, all but 30–40% of the strain rate anomalies persisted during the entire 6 years of the case study period,

    while the rest sporadically appeared or disappeared in a period from several months to a few years. The transient shifts in the

    deformation rates were remarkably synchronous with some nearby major tectonic episodes: large earthquakes and slow events.

    Differential plate coupling strengths along the subduction zones can also be inferred from the persistent pattern of rotational

    strain rate anomalies forming clockwise and counterclockwise pairs along the Pacific. Our empirical observations suggest that

    the first-order features of interseismic crustal deformations in Japan can be characterized as collateral processes behaving in

    0040-1951/$ - s

    doi:10.1016/j.tec

    T CorrespondiE-mail addr

    (2005) 27–53

    ee front matter D 2005 Elsevier B.V. All rights reserved.

    to.2005.02.003

    ng author. Tel.: +81 11 706 2643; fax: +81 11 746 7404.

    ess: [email protected] (Y. Toya).

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5328

    response to fluctuations of the tectonic stresses on multiple scales, likely influenced by changes of plate coupling strengths on

    the contiguous subduction faults.

    D 2005 Elsevier B.V. All rights reserved.

    Keywords: Active tectonic zones; Crustal deformation; Earthquake; GPS; Japan; Strain rate changes

    1. Introduction

    Tectonic boundaries are complex (e.g., King,

    1983) just as much as hierarchically distributed

    tectonic blocks (e.g., Turcotte, 1992). Particularly

    as Japan is situated at the confluence of several

    major tectonic plates (see the insert in Fig. 1, upper-

    left), tectonic block segmentation (e.g., Mogi, 1985;

    Tada, 1986) might be prevalent in the area. Between

    pairs of major tectonic blocks, there are confined

    zones of active tectonics up to a few hundred

    kilometers wide (McKenzie and Jackson, 1983) that

    often host large inland earthquakes (e.g., Jackson et

    al., 1997; Sagiya et al., 2000). A good example of

    this is the Niigata–Kobe Tectonic Line (NKTL), a

    zone of concentrated deformation that was recently

    identified using the nationwide GPS array in Japan

    (Sagiya et al., 2000). Still, the geometrical resolution

    of the suggested zone is indeterminate, and the

    nature of development or the temporal variation of

    deformation rates on such zones remains to be

    explored. Elucidating the geometry and the charac-

    teristics of deformation along dactive tectonic zonesT(Wallace, 1986) would help improve not only the

    design of the monitoring GPS networks, but also

    tectonic models (e.g., a block-and-fault model,

    Hashimoto and Jackson, 1993) hence the under-

    standing of crustal surface deformations associated

    with both intra- and inter-plate tectonics.

    The investigation of full-range tectonic processes is

    very challenging, despite the recent realization of

    nationwide continuous GPS monitoring for crustal

    deformation in Japan. The behavior of the crustal

    surface is influenced by various scale sources (in

    space, time, and magnitude): intra- and inter-plate

    interactions (e.g., Whittaker et al., 1992; Wang and

    Suyehiro, 1999; Dragert et al., 2001), readjustments of

    the crust near large seismic events (e.g., Heki et al.,

    1997; Pollitz et al., 1998; Segall et al., 2000), tides

    (e.g., Hatanaka et al., 2001; Kasahara, 2002), volcanic

    and geothermal activity (e.g., Mogi, 1958), etc.

    Arranging the monitoring GPS network in a hier-

    archical manner preferentially along active tectonic

    zones might be ideal, but it is impractical; most active

    tectonic zones have yet to be fully characterized.

    In addition, local or short-term details of crustal

    surface deformations are often obscured by the

    intricate physical makeup and local processes along

    active tectonic zones, e.g., by transrotational shearing

    (e.g., Kanaori et al., 1992; Dickinson, 1997) and by

    various transient events that do not directly reflect

    gradual and regional plate tectonics (Nur et al., 1989).

    Measurement errors in our data are also troublesome

    (e.g., Kato et al., 1998; Hatanaka et al., 2003).

    A practical approach to analyzing complex crustal

    deformation would be to focus on crustal deformation

    on one particular scale range at a time, with the

    premise that crustal deformation is a scale dependent

    phenomenon that might be seen to be operating in a

    more or less tangible unit of scale range. Such an

    approach would allow a comparison of the dynamics

    of a dlevel (scale range)T of our interest with that of theimmediate higher and lower levels in the presumable

    nearly decomposable hierarchies of crustal structures

    and processes (cf., Hierarchy Theory; e.g., Simon,

    1962; Sollins et al. (1983) in O’Neill, 1988).

    Here we are searching for inland tectonic zones in

    Japan that can be seen to be active on the mesoscale in

    the intermediate-term, and stretching the bounds of

    possibility to observe them at the maximum attainable

    resolution from the present nationwide GPS array.

    dMesoscaleT in this study refers to a dimensionapproximately twice the local mean GPS station

    separation distance and greater (approx. 70 to hundred

    kilometers). Systematic shifts in the ITRF velocity

    fields and two-dimensional instantaneous strain rate

    anomalies are regarded as active tectonic zones

    (ATZs) in this study. A set of robust and exploratory

    analyses of the continuous GPS array enables us to

    accommodate various irregularities in the data, to

    identify the sharp geometry of ATZs on the mesoscale

    nationwide, and to help decompose the multiscale

  • (b) Deceleration of upheaval(b) Deceleration of upheaval at Sakurajima V at Sakurajima Volcanoolcano

    -202

    W E

    W E

    W E

    Station ID 960719Station ID 960719Station ID 960719

    0

    S N

    S N

    S N

    0 200200200 400400400 600600600 800800800-2-2-202

    D U

    D U

    D U

    Time (days)ime (days)Time (days)

    (cm)(cm)(cm)

    -2-2-2

    2

    (Linear trend removed)(Linear trend removed)(Linear trend removed)w.r.r.t. WGS84.t. WGS84w.r.t. WGS84

    Usu VUsu Volcanoolcano Mar Mar.2000 eruption.2000 eruption

    Miyakejima VMiyakejima Volcanoolcano Jun.2000 eruption Jun.2000 eruption

    (a) Ongoing T(a) Ongoing Tokai-Kanto okai-Kanto Slow Event Slow Event

    PA

    NANA/OH/OH

    EU/AREU/AR

    PHPH

    Reference StationReference StationStations with MAD > MAD limitStations with MAD > MAD limit

    3333oN

    3030oN

    128128oE 132132oE

    3232oN

    30'30'

    3131oN

    130130oE 131131oE

    3232oN

    131131oE

    Fig. 1. Deformation velocity and acceleration fields (Example: 1998/7/18–2000/10/5). (a) Ongoing trend of Tokai-Kanto slow event. (b)

    Deceleration of an upheaval at the Sakurajima Volcano. The upper-left insert shows major/micro tectonic plates: Eurasia (EU)/Amur (AR),

    North American (NA)/Ohotsk (OH), Pacific (PA), and Philippine Sea (PH) plates.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 29

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5330

    intermediate-term processes, each of which may be

    individually related to a governing tectonic process.

    2. Method

    Careful data quality analysis of GPS array (station

    position data) in time and space makes possible the

    refinement of deformation field data on the target

    observational scale ranges: intermediate-term (~sev-

    eral months to several years) and mesoscale, respec-

    tively. Together it is important to realize that the

    observable scale ranges of crustal deformation by the

    existing GPS array are limited by the Nyquist

    sampling criteria at the lower ends (twice a unit

    sampling rate in time, provided that no higher

    frequency periodic signals are present in the data,

    and twice the local maximum station separation

    distance, given no higher frequency undulation of

    the deformation fields), and by the maximum ranges

    of the data at the higher ends (the entire duration of

    the case study period, and the whole length or width

    of the Japan archipelago). In the same way, the

    observation station arrangement plays a crucial role in

    defining the local geometrical resolvability of ATZs

    by the existing GPS array, given no high frequency

    undulation of the deformation fields.

    A temporal data quality analysis is performed at

    first, to eliminate short-term noise, such as impulse

    noise, additive seasonal variations and prominent

    transient event noise, essentially to supply only

    intermediate-term deformation rate signals to the

    subsequent spatial data analysis. The deformation

    fields as deduced from the intermediate-term signals

    are then refined to reveal only mesoscale signals by

    the application of robust smoothing and exploratory

    spatial data analysis. The keys here are to recognize

    the hidden structures and the exact type(s) of noise

    present in the deformation field data and to apply a

    suitable set of methods in handling them. As a

    result of the temporal and spatial data quality

    analyses, a reasonably uniform measure of deforma-

    tion rates is achieved nationwide on the target scale

    ranges. Using such a uniform measure of deforma-

    tion rates nationwide, the continuous monitoring of

    deformation rate changes, and the identification/

    characterization of ATZs become feasible; they help

    reveal where or which of the known tectonic (or

    seismic) zones might be currently dactiveT on themesoscale.

    2.1. Data

    The nationwide permanent GPS network, the GPS

    Earth Observation NETwork (GEONET), has been

    operated by the Geographical Survey Institute of

    Japan (GSI) since 1993 (Tada et al., 1997). Approx-

    imately 6 years of continuous GPS station position

    data on the ITRF97 (1997/6/25–2003/9/25 (UT)) were

    obtained from the GSI web-site (http://www.gsi.

    go.jp). The database was recently updated upon

    significant reduction in systematic errors and

    improvement in the database homogeneity as a whole

    (Hatanaka, 2003; Hatanaka et al., 2003). Still, system-

    atic errors (notably seasonal variations) exist in the

    data, and their mechanisms are not entirely under-

    stood. For example, regional-scale seasonal variations

    in the data are known to be well correlated with the

    seasonality of large subduction earthquakes near

    Japan (Murakami and Miyazaki, 2001), the snow

    loading cycle (Heki, 2001), etc., whereas the ampli-

    tudes of such signals are commensurate with those of

    baseline measurement errors in the array (Hatanaka,

    2003). Sudden elastic strain signals are also trouble-

    some in a study of interseismic deformation (e.g.,

    Jackson et al., 1997). In view of the above concerns,

    we attempted to eliminate prominent seasonal and

    short-term noise from the data. The culled data sets

    (presumably free of systematic errors and noise) are

    then applied in the mapping of active tectonic zones.

    2.2. Temporal data quality analyses

    2.2.1. Removal of seasonal variations in temporal data

    First, continuous daily recordings of horizontal

    station positions are selected based on the criteria

    listed in Table 1, and filtered with a five-point moving

    median to reduce impulse noise in the time series. The

    time series are then treated for removal of additive

    seasonal variations. To do this, each time series is

    carefully modeled with an equation consisting of three

    terms: seasonality (annual and semiannual compo-

    nents (e.g., Heki et al., 1997; Sagiya et al., 2000)),

    drift (a second-order polynomial), a sudden disturb-

    ance or an offset at time ts (a Heaviside function),

    besides the residuals. The timing of offset ts is

    http://www.gsi.go.jp

  • Table 1

    Criteria for selecting continuous data

    1. Data sets (stations) must belong to the permanent observation

    network. No temporary campaign data are included.

    2. Duration of data at a station must be longer than 2 years, in

    order to estimate realistic or accurate seasonal patterns. For

    this particular study, sampling time interval (dSWINT) is fixedfor 810 days (cf., Section 2.2.2).

    3. The total number of observation days (dTNODT) in a dailystation position data set must be larger than 0.75 SWIN. If either

    the data set for the East–West or North–South component of

    horizontal deformation rates at a station does not meet this

    criterion, the data for the station are not used for further

    analyses.

    4. There must be more than 2/9 TNOD for each successive SWIN/3

    interval. Those unavoidable temporal gaps in the records are

    linearly interpolated. This criterion together with the dMAD-limit(introduced in Section 2.2.2)T was effective in most cases to selectgenerally continuous data sets.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 31

    simultaneously determined while minimizing the

    least-squares error to the model, by sliding ts along

    the time-axis (with 1% increment of the whole

    duration to reduce the computation time). When the

    optimal fit is found, only the seasonal components are

    subtracted from the filtered original time-series.

    The second-order polynomial is considered for the

    drift component of the aforementioned least squares

    fitting (instead of a frequently used linear model). It is

    employed to account for the effects from sizable

    intermediate-term deformation rate changes, such as

    ones observed near the Sakurajima Volcano and the

    ongoing Tokai-Kanto slow event (e.g., ddeformationaccelerationsT, Fig. 1). At the same time, a Heavisidefunction is introduced in the model to help prevent the

    creation of artificial sinusoids at a large sudden offset

    in the data. However, the Heaviside function alone is

    not sufficient to model various types of transient

    events. Therefore, short-term irregularities, inclusive

    of earthquake related displacements, are handled in a

    different manner.

    2.2.2. Systematic removal of short-term irregularities

    Next, prominent short-term irregularities are sys-

    tematically identified and removed. A dvelocityTestimate of crustal deformation is a practical expres-

    sion for the mean deformation-rate for a given

    observation period, or a linear fit to an observed time

    series. Similarly, the second derivative of a second-

    order polynomial fit to the time series can be regarded

    as an daccelerationT of a possible nonlinear process fora well-modeled duration. Both approximations lack fit

    when large noise or transient signals are present in the

    data. Accordingly, the significance of irregularities is

    evaluated based on the goodness of fit measures of

    our approximations to the data.

    The mean absolute deviation (MAD) is used to

    measure the departure of observations from a poly-

    nomial model, per component. MAD=(A|ri�r*|)/n,where ri is the residual displacement at time i (i=1, 2,

    3, . . . , n), and r* is the sample median of the residualsafter the seasonality removal. The 96th percentile

    (approximate upper limit of the bnormal rangeQ) of thegood-fit MAD population for the case study data was

    selected as the MAD-limit: 0.25 cm for the velocities

    (Fig. 2a) and 0.2 cm for the accelerations. A MAD

    limit was defined so as to select only those data sets

    (or stations) that fit well with our models. Those

    stations with outlying MAD values above the limit

    were found clustered around concurrent transient

    events such as earthquakes and volcanic eruptions.

    The rejected stations above the limit are marked with

    square-and-cross symbols in the velocity field map,

    e.g., Fig. 1. Also, the observational time-interval is

    kept unchanged for the case studies to keep the

    measure of irregularities unbiased (810 days each [cf.,

    Table 1]). The longer the observation period, the less

    significant a transient event appears.

    The behavior of deformation-rate parameters:

    velocity and acceleration, at a transient event was

    evaluated using synthetic time series (e.g., Fig. 2b–e).

    As the models demonstrated, the MAD limit, e.g., for

    velocities (Fig. 2b, c, and e), worked effectively in

    suppressing the maximum step height of elastic

    signals (or its equivalent) in our data at ~1 cm per

    component, inclusive of the residual error amplitudes

    after the median filtering of the original time series.

    The models also suggested that some slow events

    could still be hidden in our data (Fig. 2e). A

    comparison of the spatial distributions of MAD

    between a linear model (a velocity field) and a

    nonlinear model (e.g., an acceleration field) would

    alternatively help locate such slow events, with

    preferred nonlinear fits. Although it was almost

    outside our case study period, the Bungo Channel

    slow event (Hirose et al., 1999) could be identified in

    this manner with a longer sampling time-window.

    Nevertheless, we handle the effects from slow events

  • Synthetic ModelsCase Study Data

    0 0.2 >0.4

    MAD-limit for velocity = 0.25

    Fre

    q. (

    NS

    )

    Fre

    q. (

    EW

    )

    MAD

    poor fit

    good fit

    poor fit

    good fit

    App

    aren

    t v

    eloc

    ity

    (

    cm/y

    r)

    Ste

    p h

    eigh

    t (

    cm)

    App

    aren

    t a

    ccel

    erat

    ion

    (cm

    /yr

    )2

    Maximum

    Duration of event (DE=various)

    Sampling time-interval (SWIN=810days)

    Time

    (b)

    (c)

    (d)

    Ste

    p he

    ight

    (cm

    )

    Duration of event (days)200 400 600 800

    1

    2

    3

    4

    5

    Maximum apparent velocity (cm/yr) Corresponding maximum MAD (cm)

    3

    2

    1

    1

    0.8

    0.6

    0.4

    0.2

    0.25

    MAD limit

    (e)

    TP

    (a)

    Fig. 2. bLeftN (a) Histograms of MADs concerning the horizontal velocity estimates (East–West and North–South components) using the case

    study data. Data with MADs greater than the MAD-limit (0.25 cm for velocities; cf., Section 2.2.2) are removed from the analysis. bRightN

    Behavior of deformation rate parameters at transient events on synthetic time series. (b) Synthetic displacement records of hypothetical transient

    events with a given offset-height and various durations (three time series, overlaid). (c) Corresponding apparent velocity responses around the

    hypothetical transient events shown in (b). (d) Apparent acceleration responses around the hypothetical transient events shown in (b). A time

    series record would be disturbed by a transient event for a duration dTPT=dSWINT+dDET, where dSWINT stands for the fixed sampling time-interval and dDET stands for the duration of a transient (slow) event. (e) A summary graph showing the maximum apparent velocities and thecorresponding maximum mean absolute deviations (MAD) (shown in contours) of synthetic time series as shown in (b). All models are

    evaluated using a fixed size sampling time-interval, 810 days each.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5332

    and other subduction-related phenomena differently in

    this study (discussed in Sections 2.3.3 and 2.3.4).

    2.3. Spatial data analyses

    Once the data become largely free of prominent

    short-term noise, the focus is shifted to the inves-

    Fig. 3. (a) Median, several EDA tools, and outliers. Outlier can be visu

    histogram of absolute velocity vector lengths, for example (left figure). This

    dextreme outlierT, which is greater than 3 IQR from the 75th percentileSchematic illustrations of the flow of robust smoothing process, using a de

    the upper row of the illustrations, shown are the three steps of the smoothi

    bStep IIN regularization of the observation grid and median filtering,

    independent parameters, and post-filtering. Velocity fields are depicted in

    dHNODCT indicates a nested imputation operation, described in the texoperation are: the minimum surface area, the maximum length of the si

    sampling-window radius vs. nodal-point density after the imputation oper

    point density is generally uniform when the sampling-window radius R is a

    range of F1 sigma around an estimated point density for a given R. A mapplot the graph on the left. The reason for the preferred uniform point densit

    separated by a distance of about 30 km and that they are anchored durin

    robust smoothing scheme and a fixed-point smoothing operation (moving m

    the true parameter distribution (a sine curve). Given specific ranges of (orig

    the introduced scheme can provide a nearly even overall measure of

    underestimate parameter values at local maxima and minima. (e) The ITR

    Japan; cf., Fig. 4a) for 1998/7/18–2000/10/5. The original velocity field (bl

    for comparison. Microscale irregularities of the original velocity field were

    marked by a square symbol, and the areas labeled dAT (sketched in shade

    tigation of spatial characteristics of the intermediate-

    term deformation fields. The temporal data quality

    analysis discussed in the previous section helped

    identify and remove short-term noise, whereas some

    spatial noise still remained in the data (e.g., Fig. 3a).

    Through the application of robust smoothing and

    exploratory data analysis (EDA), various properties of

    ally identified in the ITRF velocity field (right figure), and in the

    particular example of spatial impulse noise in the velocity field is an

    line in the histogram. IQR stands for the interquartile range. (b)

    formation field example with a narrow simple shear boundary. From

    ng process: bStep IN imputation of the vector field for anti-aliasing,

    and bStep IIIN representation of deformation rates in coordinate

    map views and in profiles. The illustrations are not drawn to scale.

    t (Section 2.3.2). Specific constraints applied during the HNODC

    des and the proportion of the imputation triangles. (c) A graph of

    ation (cf., Step I, b). The graph on the left illustrates that the nodal-

    bout 35 km for the case study data. A vertical error bar indicates the

    of Japan with circles on the right shows the sampling areas used to

    y around R=35 km is likely that the original observation stations are

    g the imputation operation. (d) Comparison between the introduced

    edian) using simple one-dimensional models. The top graph shows

    inal) observation station separation distances and signal wavelengths,

    the target parameter distribution. However, our scheme tends to

    F velocity fields from a small region in the case study area (central

    ack arrows) and the smoothed velocity field (grey arrows) are shown

    efficiently filtered out. (The location of the Matsushiro swarm area is

    ) indicate the general locations of tectonic boundaries.)

  • R

    RAOutlierFreq.

    25th percentile

    Median

    75th percentile

    IQR*

    >1.5 IQRaway fromthe 75th pct.

    "Outliers"~ Upper hinge

    ~ Lower hinge

    Histogram

    Boxplot

    Hinges

    e.g.) Absolute vector length

    R = Rigid blockA = Tectonic boundary

    ED

    A to

    ols

    (* IQR = Interquartile range)ITRF velocity field

    Impulse noise

    =

    ATZ ATZHNODC

    Step I: Imputation of the vector field for anti-aliasing

    Step II: Regularization of the grid and 2-D median filtering

    Step III: Representation of deformation rates in coordinate independent parameters, and post-filtering

    Query point

    Vel

    ocity

    fiel

    dV

    eloc

    ity p

    rofil

    eV

    eloc

    ity fi

    eld

    Vel

    ocity

    pro

    file

    Vel

    ocity

    fiel

    d

    Hexagonal grid for pseudo strain rate calculation

    Triangular grid from Step II, used for CMD and VLMD estimation

    ‘Impulse noise’

    ‘Edge’

    y = MEDIAN ( x1...n

    )A

    Filter window, and CMD or VLMD estimation window

    x1

    x2 xn

    x3

    (The minimum grid spacing is reflected in thecoordinate independent parameter statistics.)

    A

    (a)

    (b)

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 33

  • 0100km

    R = 35 kmR = 80 km

    R = 35 km

    20 40 60 800.02

    0.03

    0.04

    0.05

    0.06

    Radius of sampling (km)

    Den

    sity

    of p

    oint

    s (k

    m-2

    )

    Map distance

    Fixed-point operation

    Introduced scheme

    Par

    am.

    valu

    eP

    aram

    . va

    lue

    Par

    am.

    valu

    e True param. variation

    Original GPS stations

    Imputed station positions

    Predicted trend

    Velocities on original GPS stations

    Smoothed velocity field on a triangular grid

    Station 970816 w.r.t. WGS84Station 970816 w.r.t. WGS84(JUL18, 1998 ~ OCT5, 2000(JUL18, 1998 ~ OCT5, 2000

    0 400400 800800Time (days)Time (days)

    4

    2

    0

    0

    -5-5

    -10-10

    20

    -2-2-4-4

    5

    3

    MU

    D (

    cm)

    UD

    (cm

    )N

    S (

    cm)

    NS

    (cm

    )E

    W (

    cm)

    EW

    (cm

    ) Vel.= 2 cm/yrMAD = 0.1 cm

    Vel.= -4 cm/yrMAD = 0.2 cm

    Vel.= -1 cm/yrMAD = 0.4 cm

    EQs within 50 km of the station

    Spatial impulse noiseSpatial impulse noisewith clean temporal datawith clean temporal data

    (c)

    (d)

    (e) A

    Fig. 3 (continued).

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5334

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 35

    the deformation fields become increasingly clear. Our

    data generally appear to hold the following structures

    within them: dDataT=dsmooth background (brigidQtectonic blocks)T+dmesoscale edges (active tectoniczones)T+dmicroscale or impulse noises (disturbedstations, isolated tectonic features and measurement

    errors)T.EDA is an approach or philosophy of how to

    examine the data rather than a set of analysis

    techniques (NIST, 2002). The essence of EDA can

    be perceived by contrasting it to classical data

    analysis. In a classical (or confirmatory) data analysis,

    a researcher would first select a deterministic or

    probabilistic model, which consists of a set of

    assumptions for a given data set. Some quantitative

    expressions for the assumed model are then calcu-

    lated, ultimately to develop a hypothesis about the

    data (but strictly based on the model). Accordingly,

    choosing a bgoodQ set of assumptions or a ’model’ isvital for an accurate interpretation of the data. EDA,

    on the other hand, focuses on ddataT properties first,and looks for hidden data structures by utilizing

    synoptical tools such as scatter plots, histograms, box-

    plots, or any other graphical aid that might help

    visualize important features in data that might other-

    wise be missed out in a simplified deterministic model

    of a classical data analysis. As a result of EDA, a

    feasible model may be suggested by the data (NIST,

    2002). The main aim of a data analysis is to

    understand the given data, and the precision of an

    analysis result only has true significance when it is

    also accurate (Tukey, 1977).

    2.3.1. Robust and exploratory spatial data analyses

    Intermediate-term velocity fields in Japan are

    generally homogeneous and highly directional, except

    near active tectonic zones and scattered impulse noise.

    At least this is a data structure that can be immediately

    recognized in the ITRF velocity fields (e.g., Fig. 3a).

    The velocity fields on a geocentric reference frame are

    unique in the sense that they are free of the additional

    and extraneous ambiguities of locally selected refer-

    ence station(s). If there were no local crustal move-

    ments, the mesoscale velocity fields would appear

    dflatT (at least locally) and all vectors would have anearly equal length pointing in the same direction (or

    all zero vectors). For example, a test for uniformity

    (against a unimodal alternative of directional data;

    Mardia and Jupp, 2000) could be applied to report the

    highly concentrated nature of the velocity vector

    azimuths. The more concentrated the vector azimuths

    concerning a certain mean preferred orientation, the

    more rigid a local tectonic block would be, provided

    all vectors were non-zeros and had equal lengths. In

    order to describe the detailed tectonic features at our

    target scale of observation, however, it is not feasible

    to use such classical statistical analyses. The data sets

    were too sparse for the majority of our study area.

    Here instead, robust smoothing and EDA are per-

    formed to delineate mesoscale anomalies in the

    velocity fields, with added assumptions about the

    homogeneity, continuity, and the highly directional

    nature of micro- to meso-scale velocity fields. Like-

    wise, the mesoscale velocity fields are smoothed with

    the median operative as described below.

    2.3.2. Robust smoothing of the velocity fields

    The median is the 50th percentile of ordered

    sample statistics. It gives a more robust estimate of

    central location (in statistics) than the sample mean,

    and is resistant to extreme values or doutliersT(Barnett and Lewis, 1984) in data. A drobustTestimator would be insensitive to deviations from

    assumptions about a given probabilistic model or

    distribution (Huber, 1979). Meanwhile, our data (the

    velocity fields) did occasionally contain outliers

    (e.g., Fig. 3a). dAccommodationT but not manualremoval of such outliers is preferred (Barnett and

    Lewis, 1984) where it is practicable. In order to

    reduce (microscale) impulse noise but to preserve

    (mesoscale) edges in an image (the deformation

    fields), the use of a linear shift-varying filter (e.g.,

    Kalman filter) or a nonlinear filter (e.g., median

    filter) is a necessity (Huang, 1981). (This is because

    signals in two-dimensional images are always

    positive quantities and do not follow the dnormalstatisticsT (Frieden, 1979).) Therefore, taking fulladvantage of the median’s resistant attribute, the

    velocity fields are smoothed in three steps: (I)

    median imputation (filling in for missing values),

    an anti-aliasing measure for the subsequent filtering

    operations, (II) drobust smoothingT (Huber, 1979) bymedian filtering, and (III) visualization of the

    deformation fields in several alternative and conven-

    tional forms of coordinate independent parameters,

    and post-filtering. Fig. 3b illustrates the overall flow

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5336

    of the smoothing process. Most dmicroscale noisesTare polished away in steps (II) and (III), while

    dmesoscale edgesT are embossed in the dsmoothbackgroundT.

    In step (I), the velocity fields with microscale voids

    are filled in by a nested median imputation scheme,

    illustrated in the upper row of Fig. 3b. The scheme

    may be described as a hierarchical neighborhood

    operation with the median operative applied over

    constrained Delaunay triangles where the centroid of

    each triangle is the query point to be filled in

    (HNODC). The goal of this operation is to completely

    fill the case study area with nodal points as uniformly

    as possible (Fig. 3c), as an anti-aliasing measure for

    the subsequent smoothing operations. The triangles’

    proportions and surface areas are constrained to

    induce repulsion among the nodal points. Repulsion

    of these points makes their distribution much more

    regular than otherwise (Okabe et al., 1992). At the

    same time, the sample median vector-components in

    the dmarginal orderingT (Barnett, 1976) of thebivariate samples (generally independent x- and y-

    vector components) from its vertices are assigned to

    the query point at each triangle center. This median

    vector calculation, provisionally selected for its

    simplicity, gives reasonable median vector estimates,

    when the circular mean deviation of the sample vector

    azimuths is smaller than 458, i.e., when they aresamples from a highly directional velocity field. Some

    of the locally disturbed stations are also separated out

    in the process.

    In step (II), two-dimensional median filtering,

    robust smoothing is applied on the imputed velocity

    fields from step (I). This operation is effective not

    only to remove impulse noise (isolated microscale

    irregularities) but also to preserve the sharp edges of

    the robust signals in the velocity fields (active

    mesoscale tectonic zones). Justusson (1981) issued a

    detailed discussion on the statistical properties of a

    median filter. The sampling grid is simultaneously

    regularized in a triangular grid (dStep IIT in Fig. 3b).Each new grid node is assigned with a set of the

    sample medians in the marginal ordering of x- and y-

    velocity vector components: Vuv=MEDIAN [BuC],B={the nodal points in the study area from step (I)},

    and C={N nodal points inside a circular sampling

    window radius (SWR) of 35 km}. (The word

    bsampleQ is used to mean ’batch’ hereafter, since the

    observations of data after step (I) are no longer

    statistically independent.) As a result, the isolated

    noise with a surface area smaller than half of the

    sampling window (pSWR2)/2 km2 is removed. (Theedge values within an SWR from the perimeter of the

    study area are slightly extrapolated outward to abate

    probable edge effects.) More precisely, a new value on

    the regularized grid represents the majority (or NN/2)

    of the member in set C (of N nodal point values).

    However, the local nearest-neighbor distances

    among the original GPS stations (LNND) enclosing

    an imputation triangle varied regionally from 3.8 to 70

    km in our data. The Hokkaido region was particularly

    thin as for station coverage, so the LNND values were

    generally large. The frequency distribution of the

    LNND for the whole study area is unimodal,

    positively skewed and has the median and the third

    quartile of 25 km and 33 km, respectively. The

    problem of irregular station coverage was tolerably

    overcome by the anti-aliasing treatment of the velocity

    fields performed in step (I). After the robust smooth-

    ing in step (II), the nationwide distribution of

    deformation rates was uniformly rendered at a

    resolution close to the maximum attainable from the

    existing GPS array (e.g., Fig. 3d). Visual inspections

    of the actual velocity fields suggested that selecting

    the SWR of 35 km was reasonable for signals in our

    study area, as would be reevaluated later in a

    confirmatory study (Section 2.3.4). The results, e.g.,

    Fig. 3e, suffice for our exploratory objective.

    The median operations would leave a plateau

    where spatial parameter distributions were nearly

    constant. Such a plateau defines the extent of a drigidTtectonic block, and its sharp edges portray active

    tectonic zones. Along a wide active tectonic zone

    (N2SWR), the terracing of parameter variations,

    byproducts of the median operations, is anticipated.

    Still, the byproducts would have dimensions inher-

    ently smaller than 1 SWR and larger than the

    minimum size of the sampling grid-distance defined

    in step (II). So, now the idea is to filter out those

    artefacts in that range.

    In step (III), the smoothed velocity fields from step

    (II) are first transformed into several coordinate

    independent parameters. Then, additional median

    filtering with the same-size smoothing window as in

    step (II) is applied on each spatial parameter

    distribution to remove the artefacts from the previous

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 37

    step. Changing the shape of a filtering window (e.g.,

    circular or rectangular) or of the sampling grid (e.g.,

    triangular or square) did not substantially change the

    final case study results. The robust signals were

    coherent regardless.

    2.3.3. Systematic identification of active tectonic

    zones

    Anomalies in the spatial distribution of deforma-

    tion rates were expected in two general conditions: (i)

    proximity to disturbed stations and isolated micro-

    scale irregularities, and (ii) proximity to active

    tectonic zones. The condition (i) was easily noticed

    as impulses or scattered noises in the fields, when it

    did not coincide with the condition (ii). The noise (or

    outliers) in condition (i) was efficiently removed (or

    daccommodatedT; Barnett and Lewis, 1984) by therobust smoothing. Conversely, the condition (ii) was

    made prominent partly by mapping the spatial

    distribution of circular median deviation (CMD) for

    the smoothed ITRF velocity fields. CMD=MEDIAN

    (|p�|p�|hj�h*|||), where hj is the jth vector azimuth( j=1, 2, 3, . . . , n) in a sample from the smoothingwindow of step (III), h* is the sample median, and theoperands (inside the parenthesis) are the ordered

    statistics of the minimum angular distances between

    the median and sampled vector azimuths (p. 19,

    Mardia and Jupp, 2000). Also, another useful param-

    eter was the vector-length median absolute deviation

    (VLMD) for the same velocity fields. VLMD=ME-

    DIAN (|Lk�L*|), where Lk is the kth vector length(k=1, 2, 3, . . . , n) in a sample from the smoothingwindow of step (III) and L* is the sample median.

    CMD and VLMD are median deviations of the

    principal velocity vector components: azimuths and

    lengths. Both CMD and VLMD discussed here are

    specific to the ITRF velocity fields and are independ-

    ent of the orientation of the geographical coordinates.

    A comparison between CMD or VLMD distributions

    with the original velocity fields help us realize that the

    apparent bdeviationsQ do in fact represent dsystematicshiftsT in the velocity fields, or mesoscale structuralboundaries. Fig. 4 shows their spatial distributions.

    The histograms on the color-scale bars in Figs. 4a

    and b represent the frequency distributions of CMD

    and VLMD for the whole case study area. The

    overall shape of the frequency distributions is

    apparently dabsolutely outlier-proneT (Green, 1976).

    We realize, however, that despite the continuous

    appearance of the overall frequency distribution,

    there must be a nonrandom data structure within it;

    the geographical regions corresponding to the pop-

    ulations near the lower end of the overall frequency

    distribution should represent originally claimed

    drigidT tectonic blocks (R), and the other geograph-ical regions corresponding to the populations with

    longer tails under the overall frequency distribution

    should portray active tectonic zones (A). A perfectly

    drigidT tectonic block (R) and an active tectonic zone(A) are theoretically ddisjoint.T Accordingly, theoverall frequency distribution might be interpreted

    as a dcontaminated distributionT (Barnett and Lewis,1984): cja(1�k) R+kA that observations cj ofdeformation rate in a sample ( j=1,2, . . . ,n) wouldarise from a mixture of R and A , where

    0bkb1(k=contamination fraction), dRT=the popula-tion for all drigidT tectonic blocks, dAT=the contam-inants or the populations of active mesoscale tectonic

    zones.

    Nonetheless, the active tectonic zone signals (A) in

    the actual field data might include both signals: ones

    that originate from inland tectonic processes and ones

    that are directly from contiguous subduction zone

    processes with some overlaps in between. They are

    practically inseparable; they have very similar wave-

    lengths, except that the latter would be noticeable

    along the Pacific coast in Japan. Particularly, the

    geometry of mapped active tectonic zones with

    parameter values above the 75th percentile of their

    frequency distribution or the dupper hingeT (Tukey,1977) coincides well with some known tectonic

    features in the study area (Fig. 4). Accordingly,

    deformation rate anomalies with parameter values

    above the upper hinge are specifically referred as

    dATZsT thereafter. The background map in Fig. 4cillustrates a percentile summary of CMD and VLMD

    distributions, named dMD2T, highlighting the geo-graphical areas with CMD or VLMD parameter

    values greater than their 75th percentiles. In that the

    mesoscale anomalies were very powerful and con-

    centrated in confined areas, slightly shifting the center

    of the color-scale (either hue or value of the color) in

    Fig. 4 did not significantly change the overall

    appearance of the resolved ATZ distribution.

    CMD and VLMD distributions (or MD2) hint at

    the locations of ATZs. Now, two-dimensional instan-

  • Fig. 4. Maps of (a) circular median deviation (CMD) and (b) vector-length median deviation (VLMD) of the ITRF velocity fields in Japan.

    Parameter values depicted in the maps are the means of the two 810-day periods before and after the 2000 W. Tottori earthquake. The histogram

    and box-plot for each parameter are shown on the corresponding color-scale bar. The base of an arrow pointing to the right on the color scale

    indicates the upper-hinge value for the given parameter. A diamond-shaped box over the CMD map indicates the region of the velocity field

    shown in Fig. 3e. (c) Tectonic map of Japan, over a map of MD2 or a percentile summary of the CMD and VLMD distributions (cf., Section

    2.3.3). Symbols denote active volcanoes (in red triangles), Setouchi Inland Sea (in black cross stripes), and the major tectonic zones (in solid and

    dashed lines), which include: the Kuril Trench (KT), the Japan Trench (JT), the Sagami Trough (SaT), the Suruga Trench (SuT), the Nankai

    Trough (NT), the Median Tectonic Line (MTL), the Tsuruga Bay–Ise Bay Tectonic Line (TITL), the Itoigawa–Shizuoka Tectonic Line (ISTL),

    the Honjo–Matsushima Tectonic Line (HMTL) (Tada, 1986), the Niigata–Kobe Tectonic Line (NKTL) (Sagiya et al., 2000) in grey shade, and

    the Oga–Ojika Seismic Zone (OSZ) (Mogi, 1985).

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5338

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 39

    taneous strain rates are mapped to resolve the

    characteristics of intermediate-term deformations

    along individual ATZs. The smoothed velocity fields

    from step (II) are spatially differentiated to give

    coordinate independent strain rates: dilatation, max-

    imum-shear, and rotation (Terada and Miyabe, 1929;

    Tsuboi, 1930). All of these strain and rotation rate

    distributions together describe the trend of small

    strains associated with an individual ATZ. Their

    absolute quantities, however, are inexact (often under-

    estimated, cf., Fig. 3d) along ATZs and thus they are

    named dpseudo strain ratesT.

    2.3.4. Evaluation of the geometrical resolution of

    ATZs

    Before applying the introduced ATZ-mapping tool

    to the actual field data, its geometrical resolvability is

    confirmed by calibrating it against synthetic tectonic

    boundary models: (a) simple-shear and (b) pure-shear

    (or a pair of irrotational zones) models, e.g., Fig. 5.

    No dilatational strain is expected along a simple-shear

    boundary, and no rotation is expected along an

    irrotational zone. The color-scales of pseudo strain

    Fig. 5. Calibration of the introduced ATZ-mapping tool against synthetic tec

    pure-shear (a pair of irrotational boundaries) model. Dilatation (D)=(Bu(R)=[(Bu/Bx�Bv/By)2+(Bu/By+Bv/Bx)2)]0.5. No dilatational strain is expecan irrotational boundary. Three types of errors are shown as examples: a de

    On the right, two calibration test results for our study area are shown usin

    rate maps are set so as to satisfy the above require-

    ments in several synthetic models, and to highlight the

    mapped areas with absolute parameter values above

    the upper hinge of their frequency distributions using

    the case study data (cf., Fig. 4). Long-wavelength

    deformation signals and some small measurement

    errors (after step (II)) are suppressed below the upper

    hinge limit. In that most ATZ anomalies were power-

    ful on the mesoscale, the general appearance of the

    resultant maps was not substantially changed by the

    slight tuning of the color scale around the upper hinge

    in the case study data.

    It is possible to tune the overall intensity of the

    ATZ signals. For instance, a large proportion of the

    gap error, dGT in Fig. 5b, is avoided by adjusting themean nodal point density in step (I) (cf., Figs. 3b and

    c), and the sampling grid spacing in step (II). In any

    case, deflection errors, dLT in Fig. 5b, however, areinevitable within F1 SWR. Also, the widths of theATZs are adequately resolved when they are greater

    than 2 SWR, and evenly recovered when they are

    greater than twice the maximum LNND in our study

    area. Testing the mapping tool against multiple

    tonic boundary models: (a) wide simple-shear model and (b) narrow

    /Bx+Bv/By). Rotation (x)=0.5*(Bu/By�Bv/Bx). Maximum shearted along a simple-shear boundary, and no rotation is expected along

    flection error (L), a gap error or a loss of detail (G), and noises (NO).

    g the models of types (a) and (b).

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5340

    synthetic models such as those shown in Fig. 5 helps

    confirm that any powerful wide ATZs (N2 SWR) and

    non-clustered and non-oscillating narrow ATZs (b2

    SWR) are tolerably recovered by the introduced

    scheme from the present GPS array. Visual inspec-

    tions of the actual velocity fields in our study area

    suggested that the deformation rate anomalies were

    neither critically clustered nor oscillating at scales

    immediately below the mesoscale.

    The introduced scheme has an advantage over

    other fixed-point smoothing operations, e.g., a mov-

    ing median without anti-aliasing treatment on the data.

    Our method is more efficient than the others

    mentioned in illustrating the spatial parameter varia-

    tions evenly over the irregularly distributed observa-

    tion stations in our study area. This property is

    Fig. 6. Comparison between the dilatational strain rate distribution maps o

    method employed in Sagiya et al. (2000) (c, d, and e). (a) Dilatational p

    geometrical resolvability test on irrotational boundaries (cf., Section 2.3.4)

    rate map of Japan, from Sagiya et al., 2000 (reproduced, with permission, f

    (c) using the least-squares inversion method on the data for 1997/1/1–1999

    data quality analysis (cf., Section 2.2). dST-limitT or standard deviation limirregularities. The ST-limit is set to 0.33 cm (=1.3�0.25 cm) considering2003) and that the MAD-limit for velocities is set to 0.25 cm for the case st

    test on irrotational boundaries (Section 2.3.4) using the least-squares invers

    observational errors (Eq. (3), p. 2309, Sagiya et al., 2000) in the synthetic

    on the actual field data property. Dots on the maps denote GEONET stati

    demonstrated in a set of simplified one-dimensional

    models, shown in Fig. 3d. Let us suppose that the

    upper graph in Fig. 3d shows the true parameter

    distribution (a sine curve). Although the introduced

    scheme tends to underestimate parameter values at

    local maxima and minima, it can provide a nearly

    even measure of target parameter distribution, given

    specific ranges of signal wavelengths and the original

    observation station density.

    Also, our method does not employ the variance–

    covariance information of the GPS data at the

    moment; therefore, the results are not directly affected

    by the small but inevitable systematic errors in the

    variances, which can exist as error propagates in a

    large-scale network (Vanicek and Krakiwsky, 1982).

    Our results are predominantly affected by the true

    f Japan generated by the introduced method (a, b) and ones by the

    seudo strain rate map for 1997/1/1–1999/6/30. (b) The result of a

    using the same observation stations as in (a). (c) Dilatational strain

    rom n Birkhauser Verlag, Basel). (d) A reconstruction of the map in/6/30, as in Sagiya et al. (2000), after the application of the temporal

    it is applied in place of the MAD-limit to eliminate the temporal data

    that STc1.3 MAD when the distribution is bnormalQ (MathWorks,udy data (Section 2.2.2). (e) The result of a geometrical resolvability

    ion scheme, applied over the same observation stations as in (d). The

    data are assumed to be 0.13 cm for both x- and y-components, based

    ons used in the strain rate calculations.

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 41

    (immeasurable) accuracy of the available station

    position data, besides the measurement errors men-

    tioned earlier in this section.

    The improvement of the geometrical resolution of

    ATZs using the new scheme is clear, compared with

    the results by other methods, for example, ones by

    Sagiya et al. (2000). Their dilatational strain rate map

    of Japan is reproduced in Fig. 6 and compared with

    our results. They used a least-squares inversion

    scheme, a slightly modified version of the method

    introduced in Shen et al. (1996), on the GEONET data

    for January 1997–June 1999. The results by their

    method appear slightly blurred. That is likely due to

    the Gaussian spatial noise assumption in the least

    squares method and due to their disregard of the

    highly irregular observation station arrangement in the

    study area. In addition, the prominent spatial noise,

    impulse noise, does not have a clear relationship with

    the geographical distribution of temporal white noise

    amplitudes in the data. Hokkaido region in Fig. 6 was

    left blank after the temporal data quality analysis

    (Section 2.2); there would have been a risk of spatial

    aliasing affecting our results due to thin coverage of

    usable stations in the area for the specific study

    period.

    2.4. Analysis plans for spatio-temporal variations of

    deformation rates on ATZs

    The regional and temporal variations of the

    deformation fields are monitored with pseudo strain

    rates. Two types of analyses are considered: (1)

    comparisons of the spatial parameter distributions

    before and after a reference time, e.g., the time of a

    large earthquake, and (2) continuous (retrospective)

    monitoring of spatial parameter distributions. The

    observational time window was kept unchanged for

    the case study to keep the measure of deformation

    rates unbiased (cf., Section 2.2.2). For a type (1)

    analysis, the identical set of observation stations is

    used to compare the strain rate distributions for the

    two successive non-overlapping time periods (810-

    day each) before and after the 2000 Western Tottori

    earthquake. In case of a type (2) analysis, the

    observational time-window (810-day long) is moved

    along the time-axis in increments of about a month

    (30 days), and the changes in parameter strengths are

    systematically recorded at the end of the time-

    window. All available observation stations were used

    for type (2) analyses, except for data with large

    irregularities that were automatically removed from

    the analysis by the application of the MAD-limit (cf.,

    Section 2.2.2). A series of geometrical resolvability

    tests with synthetic velocity fields (cf., Section 2.3.4;

    Supplementary materials) demonstrated that not using

    the complete set of observation stations for different

    time steps did not severely affect the analysis results.

    The mean of the upper hinge values of a given

    parameter distribution for 810-day observation peri-

    ods before and after the 2000 W. Tottori earthquake

    was selected as the reference of the parameter

    strengths.

    3. Style of deformation along active tectonic zones

    Mesoscale tectonic zones on the four main islands

    of Japan were outlined in the maps of CMD, VLMD,

    and MD2 (Fig. 4), and the characteristics of current

    concentrated deformation were illustrated in the

    pseudo strain rate maps (Fig. 7). Both types of

    anomalies appeared to overlap geographically.

    Although the definition of ATZs was rather qualita-

    tive, the signals were as powerful as can be visualized

    as the systematic shifts in the ITRF velocity fields and

    as coherent strain rate anomalies. The ATZs were

    preferentially located along some known tectonic

    zones, chains of active volcanoes and low seismic

    velocity anomalies in the crust and upper mantle. The

    majority (60–70%) of the mapped ATZs was con-

    tinuously active for the whole 6-year case study

    period, while the remainder continued to operate

    sporadically.

    3.1. Dilatation

    The Japanese islands are under compression from

    the subduction of the Pacific and Philippine Sea plates

    (e.g., Kato et al., 1998). Near trench-parallel belts of

    areal contractional (negative dilatational) strain rate

    anomalies were noticed for almost the entire length of

    the island arc (Fig. 7a). The strain rate anomalies were

    persistent for the whole case study period in the

    Shikoku Island, along a belt from near Matsushiro to

    Niigata, and along the volcanic front in Tohoku and

    eastern Hokkaido.

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5342

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 43

    Contractional strain rates were exceptionally

    prevailing on the ATZs in much of Shikoku, Kinki,

    and southwestern Chubu regions before the 2000 W.

    Tottori earthquake (MJMA 7.3). In the following year,

    the Geiyo Earthquake (MJMA 6.7) also occurred 140

    km southwest of the W. Tottori event (Fig. 4a). The

    contractional strain rates at the crustal surface

    relaxed about 20% after the sequence of the earth-

    quakes (Fig. 7a). Meanwhile, a broad area from

    Kinki to Kanto appeared to be influenced by the

    accelerating Tokai-Kanto slow events (Fig. 1a) in the

    sense that it would encourage relaxation of the

    contraction in the affected upper plate areas. Sim-

    ilarly, there was an episode of intensifying areal

    contractional strain rates near the Sendai plains, a

    20% increase (Fig. 7a), before the recent sequence of

    large earthquakes in the vicinity: 2002/11/3 Miyagi-

    Oki earthquake (M6.3), 2003/5/26 Miyagi-Oki earth-

    quake (MJMA 7.1) and 2003/7/26 Northern Miyagi

    earthquake (MJMA 6.4).

    Positive dilatation was remarkable near the Sakura-

    jima Volcano, e.g., for 1998/7/18–2000/10/5 (Fig. 7a),

    and the deceleration of an upheaval was evident in the

    acceleration field (Fig. 1b). An acceleration field

    associated with a permanent deformation, which

    would be proportional to a set of driving forces in a

    nearly uniform medium, is an interesting subject to

    study in detail. However, it is beyond the scope of this

    paper, and will be explored elsewhere.

    3.2. Shear

    Fig. 7b shows the distributions of maximum shear

    strain rates. Large shear strain rates were detected in

    several areas: over much of Shikoku and the

    southeastern Kyushu islands, near the Atotsugawa

    fault, in the Matsushiro–Niigata area or the northern

    Fig. 7. Pseudo strain rate maps: (a) dilatation, (b) maximum shear, and (c

    (Period 1), and the bottom one is for 2000/10/6–2002/12/25 (Period 2). Th

    value of a given absolute parameter frequency distribution. Temporal variat

    in two selected areas: Zone 1 and Zone 2, are graphed in the inserts for (a)

    illustrate the fluctuations of the sample medians of maximum shear pseudo

    line illustrate the fluctuations of the corresponding IQRs. IQR stands for t

    records (cf., strain rate calculation points, Fig. 3b), and the time series o

    Section 2.2.2) is changing with time. Labels in the inserts for (b) denote: 20

    earthquake (M1), 2003/5/26 Miyagi–Oki M7.1 earthquake (M2), and 200

    denote: Shimanto metamorphic belt (Shimanto), Abukuma metamorphic t

    (data sources: Japan Meteorological Agency and their affiliates (2003), an

    NKTL, near the Sagami Trough, along the volcanic

    front in Tohoku (Nakajima et al., 2001) and eastern

    Hokkaido, and above a low seismic velocity

    anomaly in the Hidaka range (Takanami, 1982;

    Murai et al., 2003).

    Shikoku and southern Kyushu are extensively

    strained in the intermediate-term, but with very little

    MN3 seismicity. Particularly, the Nankai forearc sliver

    to the south of Setouchi or the Median Tectonic Line

    (MTL) is in a fast right-lateral shear with respect to

    the dRigidT Chugoku block (Fig. 8) and internallybrotatingQ (Fig. 7c; cf., Section 4.3). Fig. 8 showsMTL-parallel and -normal components of deformation

    rates in the Shikoku Island with respect to the

    Chugoku block. The apparent aseismic movement

    on MTL is roughly 5 mm/year right-lateral, which

    agreed with the rate estimated by others (Miyazaki

    and Heki, 2001; Tabei et al., 2003). What is striking is

    that both MTL-parallel and -normal deformation-rates

    on the Shikoku Island decreased several millimeters

    per year after the 2000 W. Tottori earthquake (upper-

    right graphs in Fig. 8). This observation can be

    compared to the aforementioned decrease of areal

    contractional strain rates around the turn of the

    century (Fig. 7a). Both components of the deforma-

    tion rates along the MTL contained parts of the

    trench-normal strain rate changes, because of the

    obliquity of the MTL with respect to the Nankai

    Trough (lower-left map in Fig. 8).

    Southwestern and northeastern Japan are in colli-

    sion at central Japan on a plate boundary between the

    Amurian and North American plates (Miyazaki and

    Heki, 2001). The rate of convergence is approxi-

    mately 2 cm/year in the EW direction, derived by

    estimating the relative drigidT block movements acrossan ATZ near the junction of the Itoigawa–Shizuoka

    Tectonic Line (ISTL) and the NKTL (Fig. 9). This

    ) rotation. The upper map of each pair is for 1998/7/18–2000/10/05

    e base of an arrow drawn on a color-scale indicates the upper-hinge

    ions of dilatational and maximum shear pseudo strain rate parameters

    and (b) on the right. Thick lines in the time series (the insert for (b))

    strain rates in Zones 1 and 2, and two thin lines paralleling a thick

    he interquartile range. dNT stands for the number of grid nodes withf dNT illustrate how the density of usable observation stations (cf.,00/10/06 W. Tottori earthquake (WT), 2002/11/03 Miyagi–Oki M6.3

    3/7/26 Northern Miyagi M6.4 earthquake (M3). Labels on map (c)

    errane (Abukuma), and Hidaka range (Hidaka). (d) Seismicity map

    d Utsu (1982)).

  • Maximum Shear Pseudo Strain Rate

    0.5

    1.5

    1

    2x 10

    -7

    0

    (Average of Period 1 and 2)MTL normal

    MTL parallel

    0 M

    TL

    0 M

    TL

    AA

    NTNT

    3

    2

    1

    0

    MTL normalMTL normal

    Con

    trac

    tiona

    l v

    eloc

    ity w

    .r.t.

    'A'

    (

    cm/y

    r)

    50 100 150 200 3

    2

    1

    0

    MTL parallelMTL parallel

    Distance from 'A' kmW

    este

    rly

    vel

    ocity

    w.r.

    t. 'A

    '

    (cm

    /yr)

    decrease { Vel.(Period2) - Vel.(Period1) > 0 }increase { Vel.(Period2) - Vel.(Period1) < 0 }

    MTL MTL A A

    0

    Deformation Rate:

    a few mm/yr

    MTLMTLForearcForearc SliverSliver

    r

    MTL

    6.56.5 cm/m/y

    NT

    Cont. 3x10Ext.

    -7Principal Strain Axes

    Shikoku

    (Convergence rate, Miyazaki and Heki, 2001)

    /yr

    /yr

    Fig. 8. Crustal deformation in Shikoku region. Shown are Median Tectonic Line (MTL) parallel and normal components of the velocity fields

    with respect to the drigidT Chugoku block. The group of stations in the box with cross-stripes along A–A line was considered as the fixedreference. On the right, two sets of velocity field profiles are shown. Period 1=1998/7/18–2000/10/05. Period 2=2000/10/06–2002/12/25. Two

    things can be noted. (1) Small right-lateral slip along the MTL was apparent, 5 mm/year at the most. (2) There was an orderly decrease in the

    deformation-rates (both MTL parallel and normal components) on the forearc in Shikoku around the turn of the century.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5344

    result agrees with the estimate by others (Miyazaki

    and Heki, 2001; Hirahara et al., 2003), determined by

    other means. Fast creep movements along the

    Atotsugawa fault on the NKTL appeared to be driven

    by such a fast convergence in the area. When the

    relative velocity vector was projected onto an NKTL

    segment near the Atotsugawa fault with a strike of

    N60E, a rapid right-lateral shear (1.6 cm/year) and a

    fault-normal contraction (1.3 cm/year) were detected

    for 1998/7/18–2000/10/5, for example. The median

    absolute deviation for both components was ~0.35

    cm/year. GSI (1997), otherwise, reported local varia-

    tions of creep rates along the central portion of the

    Atotsugawa fault: 1.5 cm/year for a fast moving

    segment and non-creeping at a position 15 km west of

    it. These estimates were based on multiple precise

    distance measurements with several 1–2 km baselines

    across the fault. The last large earthquake to occur in

    the area was the 1858 Hietsu earthquake (M7)

    (Usami, 1996).

    3.3. Rotation

    Rotational ATZs were noticeable along the NKTL

    (clockwise) and along the Pacific. Several clockwise

    and counterclockwise pairs of rotational ATZs were

    perceived in areas along the Pacific: the Shimanto

    metamorphic belt (southern Kyushu to Shikoku-Kii),

    Tokai, Sagami, the Abukuma metamorphic terrane to

    the Kitakami mountains, and the Hidaka range to the

    volcanic front in eastern Hokkaido (Fig. 7c). The

    direction of maximum compressive stress, as deduced

    from the focal mechanisms of shallow inland earth-

    quakes, is known to be the same as that of the plate

    convergence (E–W) in northeastern Japan, while it is

    near trench-parallel (SW–NE) (Wang and Suyehiro,

    1999) and occurs at an angle of about 458 to the plateconvergence direction in southwestern Japan (McCaf-

    frey, 1993). Supposing that an in-between block of a

    rotational ATZ pair was compressed faster than the

    surrounding areas in the fashion depicted in the insert

  • Fig. 9. Pseudo strain rate maps of central Japan, for 1998/7/18–2000/10/05. Shown are, from the left, dilatational (D), rotational (x), andmaximum-shear (R) components. The base of an arrow drawn on a color-scale indicates the upper-hinge value of a given absolute parameterdistribution. The spatial distributions of the dilatational and maximum shear strain rates across two anomalies are also shown in profiles. A red

    line on the profiles indicates the upper hinge value for the corresponding parameter. The active faults are shown in grey lines on the maps (active

    fault data source: Research Group for Active Faults of Japan, 1991).

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 45

    of Fig. 7c, then, the sense of rotation on the ATZ to

    the right of the in-between block facing the compres-

    sion direction would rotate clockwise and that of the

    left would rotate in a counterclockwise direction.

    Given such a configuration, it is possible to recognize

    that landward crustal movements of the in-between

    blocks were prevailing during the whole 6-year

    observation period, except for a pair in Tokai that

    indicated a seaward transient crustal movement. The

    trend and timing of the crustal surface movement in

    Tokai (Fig. 1a) matched that of the ongoing Tokai

    slow earthquake (Ozawa et al., 2002).

    4. Discussion

    Our empirical observations, refined by the appli-

    cations of robust smoothing and exploratory–confir-

    matory analyses on the irregular data sets,

    demonstrated that the primary features of the inter-

    seismic crustal deformation in Japan can be described

    in terms of two overlapping operative processes: (1)

    the regional-and-steady mode of strain accumulation

    (cumulative effect of distant tectonic forces) and (2)

    the local-and-transient mode of strain accumulation

    (individually linked to an imminent or concurrent

    sequence of large tectonic events, both inland and

    offshore).

    Mesoscale intermediate-term deformation signals

    (here altogether recognized as of ATZs) often

    account for slow earthquakes (e.g., Hirose et al.,

    1999), post-seismic deformations (e.g., Heki et al.,

    1997), elastic strain accumulations and permanent

    internal deformations in the upperplate areas. ATZs

    accommodate interseismic deformation of various

    forms: elastic strain accumulation (Jackson et al.,

    1997), creep along a confined zone of crustal

    weakness (e.g., GSI, 1997), and wall-block perma-

    nent deformation (Sibson, 2002). Superposed upon

    these localized inland crustal movements, there are

    broader tectonic loading signals of various scales

    from the contiguous major plate boundaries in

    Japan. Widespread disagreements between regional

    instantaneous geodetic strain rates (�10�7/year)

  • In space: MESOSCALE (70~several hundred kilometers)In time: INTERMEDIATE TERM (several months to years)

    Our field of view / Observational scale range

    Operational scales of interseismic crustal deformation

    SPACE

    long(several-100s years)

    intermediate (months-several years)

    short(seconds-months)

    large(100s-1000s km)

    medium(10s-100s km)

    small(0-10s km)

    Our field of view

    Static stress-change

    effect dominant

    TIME

    Regional-and-steady processLocal-and-transient process

    Operative / characteristic scales of crustal deformation in our field of view:

    Highly heterogeneous

    mixture of effects + noise

    Dynamic stress-change

    effect dominant

    Fig. 10. Summary diagram of the operational scales of interseismic

    crustal deformation with respect to the observable range by the

    existing GPS array. Our observable scale-range is limited within the

    rectangular box in the center, labeled dOur Field of ViewT. Thediagram would help distinguish the regional-and-steady deforma-

    tion signals (of scales around the solid circle) from local-and-

    transient deformation signals (of scales around the open circle).

    Local-and-transient strain rate changes are presumably caused by

    stress-changes accompanying the bco-seismicQ phenomena prevail-ing on certain scale-ranges, labeled in the background. bCo-seismicQin this context might include immediate precursory deformation and

    immediate post-seismic deformation (both fast and slow).

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5346

    (Kato et al., 1998; Sagiya et al., 2000) and local

    geologic strain rate estimates along active fault

    traces (�10�8/year) (Nohara et al., 2000) are oftenattributed to such interseismic elastic strain accumu-

    lation due to broad tectonic loading from the

    contiguous subduction zones, where only small

    portions of the differences actually contribute to

    the permanent strains on many individual faults.

    (Additionally, the fact that not all active microscale

    or blind faults are identified in the crust might

    explain a part of the large disagreements.)

    Subduction effects (on both meso- and macro-

    scales) should be removed from our data to correctly

    assess the upperplate and intraplate deformations

    (personal communication with several at IUGG,

    2003). However, separating these effects presents

    quite a challenge, as Japan is situated at the

    confluence of several major tectonic plates. If one

    were to identify and remove various subduction

    effects, it would be necessary to perform the tasks

    comprehensively; otherwise, it could merely create

    biases in one’s interpretation of inland tectonics. The

    maps of coordinate independent deformation-rate

    parameters (Figs. 4 and 7) alternatively helped

    identify mesoscale areas of disturbances or ATZs in

    the upperplate that likely resulted from both intra- and

    inter-plate tectonics. The resolved geometry and

    characteristics of the ATZs might be utilized in a

    tectonic model (e.g., Hashimoto and Jackson, 1993)

    and could provide an improved understanding of

    heterogeneous and non-isotropic mesoscale tectonic

    processes in the near future.

    4.1. Geometry and origin of inland ATZs

    The geometrical agreements among the mapped

    ATZs, chains of active volcanoes, and low seismic

    velocity anomalies in the lithosphere are apparently

    due to their common weakness on the mesoscale.

    Moderate to large inland earthquakes (M5.7–8.0,

    Db=20 km, in Japan (Zhao et al., 2000)) also tend

    to occur near such preexisting weaknesses in the

    crust (Zhao et al., 2000; Sagiya et al., 2000). The

    weakness is believed to come from a high water

    content in the shallow crust originating from the

    dehydration of the subducted slabs (Nakajima et al.,

    2001; Iio et al., 2002; Hyodo and Hirahara, 2003).

    The presence of partial melt materials in the upper-

    most mantle, e.g., beneath the volcanic front in

    Tohoku (Nakajima et al., 2001), might be another

    contributing factor for the relative structural weak-

    ness on the mesoscale. Leaving aside the issue of

    whether these weak zones constitute micro-plate

    boundaries, they were apparently more sensitive to

    ambient stress changes than the neighboring drigidTblocks. The apparent weakness on the mesoscale,

    however, does not necessarily imply the same

    quality for dmicroscaleT ATZs. Our observations arelimited to the scale-range in our field of view (Fig.

    10). Similarly, the apparent behavior of crustal

    surface materials whether it is ductile or brittle

    critically depends on the observational scale (e.g.,

    King, 1983).

    A focused strain rate anomaly with its instability

    apparent in a certain geometric condition of tectonic

    boundaries, e.g., a triple junction on land (King,

    1983; Gabrielov et al., 1996), might imply a stress

  • Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 47

    concentration. For example, the 1965–1967 Mat-

    sushiro earthquake swarm area appeared to be

    located near a junction of three mesoscale ATZs

    (Figs. 3e and 4a). With southwestern Japan in

    collision against the northeastern half (Fig. 9), the

    localized stress near Matsushiro was possibly close

    to its critical state in 1965 and eventually developed

    into an energetic ductile deformation. The source of

    locally dexcess (continuously been replenished)Tseismic energy-inputs during the swarm (Kisslinger,

    1968) might have been the momentum in such a

    ductile deformation. Expansive strains in the NS

    direction were distinctive of the swarm episode,

    which also matched well with the concurrent

    moderate seismic events’ mechanisms (Kasahara

    and Okada, 1966). It is likely that water eruptions

    and other unusual phenomena were secondary to

    such a diffuse process. In addition, the 1964 Niigata

    earthquake (224 km north–north–west of Matsush-

    iro) appeared to trigger the Matsushiro swarms, as

    witnessed in the migrating pattern of diffused

    seismicity from Niigata to Matsushiro (Mogi,

    1988). In general, prominent ATZs, such as those

    near Matsushiro and Akan in Fig. 4a, are suggestive

    of well-developed tectonic zones, highly strained

    and heterogeneous areas, where continued readjust-

    ments and ductile deformation might possibly be

    taking place on the mesoscale. Further investigations

    of the swarm areas might be necessary to confirm

    the above conjecture.

    4.2. Subduction-induced strains on ATZs

    Tectonic stresses in the overriding plate of a

    subduction zone in the absence of back-arc spreading

    are critically influenced by the locking and unlocking

    of the adjacent subduction fault (Whittaker et al.,

    1992), and the stresses fluctuate with time over many

    earthquake cycles because of such a locking-and-

    unlocking mechanism (e.g., Wang and He, 1999).

    Temporal fluctuations of the stresses in turn induce

    strain rate changes in the overriding plate that may be

    observed by geodetic means. In our 6-year GPS

    observation period, which was much shorter than the

    interseismic period of a large characteristic earthquake

    such as the Nankai earthquake, we still perceived at

    least two apparent modes of strain accumulation along

    the island arc. This is illustrated in the summary

    diagram of operational scales of interseismic crustal

    deformation, Fig. 10.

    (1) Regional-and-steady mode (scales around a

    solid circle in Fig. 10). About 60–70% of the

    mapped inland ATZs or strain rate anomalies

    were perceptible for the whole 6-year observa-

    tion period; the overall pattern (the distribution

    and the trend) of the anomalies did not

    significantly change over time nationwide.

    These steady strain rates presumably reflect

    the gradual interseismic loading from the

    adjacent subduction zones.

    (2) Local-and-transient mode (scales around an

    open circle in Fig. 10). The remaining part of

    the anomalies occurred only for brief moments,

    from several months to a few years, and affected

    zones from 70 to a few hundred kilometers

    wide. For example, there were two notable cases

    where the transient shifts in areal contractional

    and shear strain rates were remarkably synchro-

    nous with the sequences of nearby major

    tectonic episodes including both large earth-

    quakes and slow events (Figs. 7a and b;

    Sections 3.1 and 3.2).

    In Shikoku and Kinki regions, the areal contrac-

    tional strain rates had weakened suddenly at the turn

    of the century, when the W. Tottori and Geiyo

    earthquakes occurred (Figs. 7a and 8). The Tokai

    slow event (Ozawa et al., 2002) (or aseismic slip on

    the subduction fault) was also happening some 500

    km away east of the W. Tottori epicenter. All of these

    events would cause the stresses in the upper plate

    areas to decrease.

    In order to determine the stress changes caused by

    transient aseismic slips on a subduction fault, it is

    essential to have precise knowledge of the spatial

    extent, magnitude, and duration of the slips. However,

    it is quite difficult to estimate these items solely from

    an observation of crustal surface deformation. The

    most useful vertical component of strain has the

    poorest resolution with GPS observations (e.g.,

    Melbourne and Webb, 2003).

    Optionally, there is a possibility that transient rate

    changes in microseismicity (MN=3) on the subduction

    fault (a lower right graph, Fig. 11) might help

    determine the precise duration and extent of the

  • Inland Seismicity

    Subduction Slab Seismicity

    4

    6

    4

    6M

    M

    048

    12x 10

    20

    Cum

    ulat

    ive

    en

    ergy

    rel

    ease

    7 x 10(mainshock)

    21

    1930

    R=25km

    R=150km

    Quiescence

    1940 1950 1960 1970 1980 1990 2000

    R=25km

    Inland Seismicity before W.Tottori Earthquake (Depth < 25km)

    1998 1999 2000 2001 20020

    0.2

    0.4

    0.6

    0.8

    1

    Time [year]

    Cum

    ulat

    ive

    num

    ber

    (nor

    mal

    ised

    )

    Subduction Slab Seismicity (M>3)

    WT (inland)

    GY (intra-slab)

    Zone E

    Zone W

    130 132 134 136 13831

    32

    3334

    35

    36

    R=25km R=150km

    Zone E (N=964)

    Zone W (N=1420)

    GY (Depth=51.4km)

    WT (Depth=11.3km)

    Fig. 11. Seismicity rate changes around the turn of the century in southwestern Japan. (1) Inland seismicity before the 2000 W. Tottori

    Earthquake, within 25 km of the epicenter (top), and within 150 km of the epicenter (bottom). (2) Subduction slab seismicity. WT=W. Tottori

    earthquake. GY=Geiyo earthquake. The microseismicity rate in Zone E changed at 2000 (increased out of the background level), at dWTT(further rate increase, indicative of a triggered aseismic slip), and at 2001 (decreased back to the background level). The earthquake catalogue of

    the Japan Meteorological Agency and their affiliates (2003) used here was generally dhomogeneous (Habermann, 1983)T and dcomplete (e.g.,Wiemer and Zuniga, 1994)T.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–5348

    aseismic slip in combination with geodetic observa-

    tions. A separate detailed study on this subject is

    necessary. (It would also be interesting to investigate

    the relationship between the non-volcanic tremors

    (Obara, 2002) and slow events, as was done in the

    Cascadia subduction zone (Rogers and Dragert,

    2003).) Since the shear stress changes on the fault

    would likely be reflected as changes in the micro-

    seismicity rate, it is plausible that the plate coupling

    strengths on the Tokai to Tonankai (and possibly a

    part of Nankai) subduction faults had weakened

    slightly during 2000 to 2001, followed by the Geiyo

    earthquake in 2001 (Fig. 11). Particularly, the flow of

    these tectonic events appeared in sync with the

    decrease of areal contractional strain rates on the

    overriding plate regions (Figs. 7a and 8). It is

    conceivable that the scale of crustal deformation and

    seismicity would correlate well only when their

    operative scales match.

    The situation in Miyagi Prefecture was similar to

    the case in southwestern Japan insofar as the areal

    contractional and shear strain rates were higher in the

    overriding plate before the recent sequence of large

    earthquakes (both inland and offshore events). The

    tectonic setting in northeastern Japan is overall differ-

    ent from that of southwestern Japan. The overriding

    plate in northeastern Japan is under trench-normal

    compression and the rate of subduction is much faster

    (and the slab is colder) than that of southwestern Japan

    (Wang and Suyehiro, 1999). Accordingly, the earth-

    quake production rate in northeastern Japan is much

    faster than that of the southwest, for example.

    The strain is evidently being accumulated near

    Miyagi Prefecture. The Headquarters for Earthquake

  • ID 950392

    ID 950393

    WT

    Station ID: 950393; on WGS84 (annual and semi-annual additive seasonality and linear trend removed)

    0 400 800 1200 1600

    NS(cm)

    Time (days)

    -1

    0

    1 WT

    Fig. 12. Small precursory displacements detected at a few GEONET

    stations (e.g., station d950393T) near the 2000 W. Tottori earthquakeepicenter (dWTT). Station position time series at the station d950393T(North–South component) on the WGS84 reference ellipsoid is

    shown. Data are for 1998/7/18–2002/12/25. Signals at the station

    d950392T were also very similar to those at the station d950393T.dWTT=W. Tottori earthquake.

    Y. Toya, M. Kasahara / Tectonophysics 400 (2005) 27–53 49

    Research Promotion in Japan (ERP at http://

    www.jishin.go.jp) gave their forecast for the next

    M7.5-class interplate Miyagi-Oki earthquake to

    occur in the next 30 years at the probability of

    90% or greater, as estimated from earthquake

    recurrence-time studies.

    Wang (1995), for example, viewed that strain

    accumulation at a subduction zone might take place

    on multiple spatial scales. Elastic strain energy

    induced by distant tectonic forces is stored up in a

    broad region in the lithosphere, while the strain is also

    accumulated locally around isolated asperities on a

    subduction fault. Locally accumulated strain is even-

    tually released as earthquakes near asperities. Because

    fast crustal shortening (inelastic permanent deforma-

    tion) in the overriding plate is typically not observed,

    the remaining stored strain energy must be released by

    some other effective mechanism(s) conceivably on the

    subduction fault (Wang, 1995). With the knowledge

    of prevalent low seismic coupling factors at world-

    wide subduction zones, compiled by Pacheco et al.

    (1993), Wang (1995) suggested that aseismic slip

    must be the predominant mechanism of strain releases

    at a subduction zone, where the growth of a single

    seismic slip with a rapid elastic rebound of the

    overriding plate is hindered by the viscous force in

    the asthenosphere. Although the viscoelasticity of the

    asthenosphere may not be the sole factor in control-

    ling the low seismic coupling factors at world’s

    subduction zones (Hirahara, 2002), our observations

    are not in conflict with Wang (1995)’s observations of

    a common subduction zone environment where

    aseismic slips are prevalent and multiscale strain

    accumulation zones exist. If each pair of rotational

    ATZs along the Pacific (Fig. 7c) is presumed to

    constitute a subduction-induced strain accumulation

    zone, some of the pairs would make up zones several

    hundred kilometers wide. They are clearly larger than

    a single historic interplate earthquake rupture in the

    area. The prevalence of aseismic slips and the

    existence of nested strain accumulation zones seem

    to be in accord with our findings.

    4.3. Limitations of the scheme

    Our ATZ-mapping tool successfully delineated

    sharp mesoscale tectonic boundaries; however,

    dmicroscaleT features were outside our perceivable

    range of scale (Fig. 10). Therefore, we were unable to

    deduce the style of microscale deformations, e.g.,

    along a rotational ATZ, whether they involved multi-

    ple distinct block rotations or more diffused deforma-

    tions within a zone of concentrated deformation. Also,

    most precursory deformations of an inland earthquake

    (if there were any tangible ones) on an imminent

    rupture surface,